Geographical and Energy Aware Routing:
a recursive data dissemination protocol for wireless sensor networks
Yan Yu, Ramesh Govindan, Deborah Estrin
Email: yanyu/[email protected], [email protected]
08/14/2001
Abstract
Future sensor networks will be composed of a large number of densely deployed sensors/actuators. A key feature of
such networks is that their nodes are untethered and unattended. Consequently, energy efficiency is an important design consideration for these networks. Motivated by the fact
that sensor network queries may often be geographical, we
design and evaluate an energy efficient routing algorithm
that propagates a query to the appropriate geographical region, without flooding. The proposed Geographic and Energy Aware Routing (GEAR) algorithm uses energy aware
neighbor selection to route a packet towards the target region
and Recursive Geographic Forwarding or Restricted Flooding algorithm to disseminate the packet inside the destination region.
We evaluate the GEAR protocol using simulation. We find
that, especially for non-uniform traffic distribution, GEAR
exhibits noticeably longer network lifetime than non-energyaware geographic routing algorithms.
1 Introduction
Future sensor networks will be composed of a large number of densely deployed sensor nodes. Each node in the sensor network may consist of one or more sensors, a low power
radio, portable power supply, and possibly localization hardware, such as a GPS (Global Positioning System) unit or a
ranging device. A key feature of such networks is that their
nodes are untethered and unattended. Consequently, they
have limited and non-replenishable energy resources. Therefore, energy efficiency is an important design consideration
for these networks.
In this paper we study energy efficient geographic packet
forwarding techniques. Disseminating information to a geographic region is a very useful primitive in many locationaware systems, and especially sensor networks. For example, a sensor net application may be interested in “what is
the average temperature in a region R in the time period
(t1 ; t2 )” [10]. The region can be expressed, for example,
by a rectangle in 2-space. In order to fulfill the above communication task, this query needs to be disseminated to the
sensors in the specified region.
An efficient way to disseminate the geographic query to a
specified region is to leverage the location knowledge in the
query and to route the query directly to the region instead of
flooding it everywhere. Previous research has studied how to
geographically route a packet to a target location in an ad-hoc
network [12]. However, our work compares with previous
geographic routing as follows:
1. Unlike unicast communication in previous systems, we
study the problem of forwarding a packet to all the
nodes inside a target region, which is a common primitive in data-centric sensor net applications [10].
2. Furthermore, our work does not assume the need for
a location database that maps node identifier to node
location. We expect sensor networks to be data-centric,
where communication primitives are expressed not in
terms of node identifiers but in terms of named data.
3. Our work also assumes static (i.e., immobile) sensors.
This does not simplify the geographic routing problem,
but does enable some of the route learning techniques
we use.
4. Like previous work, however, we do assume the existence of a localization system [22, 2, 18, 7, 19] that enables each node to know its current position.
5. Motivated by the stringent energy constraint in sensor networks, we use energy aware metrics, together
with geographical information, to make energy efficient
routing decisions. In previous work, balancing energy
usage has not been an explicit design goal.
Our Geographic and Energy Aware Routing (GEAR)
technique uses energy aware and geographically informed
neighbor selection heuristics to route a packet towards the
target region. Within a region, it uses a recursive geographic
forwarding technique to disseminate the packet. Although
the energy balancing design of GEAR is motivated by sensor net applications, our protocol is generally applicable to
ad-hoc networks.
We simulated GEAR for uniform and non-uniform traffic
distributions, and compared its performance to GPSR [12],
which is a non energy-aware geographic routing algorithm.
For non-uniform traffic, GEAR delivers 70% to 80% more
packets than GPSR. For uniform traffic, GEAR successfully
delivers between 25% and 35% more packets than GPSR.
However, in both cases, GEAR performs significantly better
in terms of connectivity after partition—the fraction of pairs
remaining connected after a “partition” (when all sources
are partitioned from their respective target regions). We are
currently implementing a prototype of GEAR protocol in a
moderate size testbed.
The rest of the paper is organized as follows. We briefly
discuss related work in Section 2. We explain GEAR in detail in Section 3, and present our simulation results in Section 4. We discuss some design details and describe the ongoing implementation work in Section 5. Section 6 presents
our conclusions.
on the destination’s previous known location and its known
mobility pattern. However, when mobility information is not
accurate, the request zone may have to be enlarged to the
whole network.
Li et al [15] propose a scalable and distributed location
database service, which tracks mobile nodes’ locations. It
selects multiple location servers to store each node’s location. Queries for a mobile node’s location are resolved using the predefined identifier ordering and spatial hierarchy
to find a location server for that node. As mentioned before,
such a location database service is not necessary in our target
application.
2 Related work
2.2
2.1
Geographic ad-hoc routing
Most previous geographic routing protocols use greedy
algorithms to forward the packet to the destination. They
differ in how they handle communication holes.
Finn [4] is the earliest work known in geographical routing. He used restricted flooding search to navigate around
holes. One drawback of this mechanism is the difficulty in
determining an appropriate scope for the search. GPSR, by
Karp et al [12], elegantly avoids this problem by deriving a
planar graph out of the original network graph. In GPSR, the
packet follows the perimeter of the planar graph to circumvent holes. The derived planar graph is much sparser than
the original one, and the traffic concentrates on the perimeter of the planar graph in perimeter mode. Thus, the nodes
on the planar graph tend to be depleted quickly. In addition,
nodes are assumed to operate in promiscuous listening mode
and consequently consume energy [21].
Scalable Location Update-based Routing Protocol(SLURP) [24] constantly maintains approximate location
information of nodes in the network, and finds accurate
routes to specific nodes on demand. It uses approximate
geographic routing to route a packet to the region that
contains the destination, and once the packet is inside that
region, it uses source routing to reach the destination. It
relies on route request to circumvent holes. The route
request/ reply overhead and constant snooping mode in
SLURP make it unsuitable for sensor net applications.
Less directly relevant is the work of Imielinski and
Goel [9], who propose querying and monitoring DataSpace.
One primitive in this application is to send a query to a datacube, which is analogous to forwarding a query to a certain region in this paper. They first use geographic routing
to forward the query to the geonode responsible for the datacube specified in the query. Then the corresponding geonode multicasts the query to relevant nodes based on its prebuilt indexing structure. However, it is difficult to build and
maintain an efficient index structure under the high level of
dynamics that sensor networks are exposed to.
Ko et al. propose Location Aided Routing (LAR) [13],
which limits the search for a new route to an estimated “request zone”. The geographic location information is not used
to make routing decision, but to limit the route request flooding to a smaller region. The request zone is estimated based
Other related work
Ad-hoc routing Dynamic Source Routing (DSR) [11]
floods route request on-demand, and suggests using promiscuous listening to aggressively learn new routes. Sensor networks have more stringent energy constraint than do ad-hoc
networks, and bandwidth requirements will be lower in sensor networks. Therefore, flooding route request or operating
in promiscuous mode make DSR undesirable for sensor networks. Other existing Ad-hoc protocols [17, 16] also require
greater energy resources and higher bandwidth demand than
is anticipated for sensor networks. However, some of our
techniques might be applicable in energy efficient versions
of these protocols.
Energy aware routing Rather than using traditional metrics such as hop-count or delay for finding routes, Woo et
al. [23] proposes five energy aware metrics such as “maximize time to partition” and “minimize maximum node cost”.
These are important metrics for energy efficient routing,
however, it is difficult to directly implement them in a local
algorithm when even the global version of the same problem
is NP-complete.
Chang et al. [3] proposed a class of flow augmentation
algorithms and a flow redirection algorithm which balance
the energy consumption rates among the nodes in proportion
to the energy reserves. The limitation of this approach is
that it requires the prior knowledge of the sets of origin and
destination nodes and the information generation rates at the
origin nodes, consequently, the topology and the traffic are
fixed at least between consecutive computations.
LEACH [8] proposes a clustering based protocol that utilizes randomized rotation of local cluster heads to evenly distribute the energy load among the sensors in the network. It
is similar in spirit with other energy aware routing protocols in terms of load balancing. However, their underlying
assumption is different from ours. They assume adjustable
transmitting power and assume that the cluster head talks
directly to the gateway node. We assume each node has
fixed transmission power – optimizing transmission power
in a multi-hop network is outside the scope of this paper.
PAMAS [20] proposes a new power aware multi-access
protocol for ad-hoc radio networks. It conserves battery
power at nodes by intelligently powering off nodes that are
not actively transmitting or receiving packets. Their energy
efficient MAC design complements our work.
Sensor network routing mechanism Directed diffusion [10] is a data-centric protocol for sensor network applications. It achieves some level of energy savings by selecting empirically good paths, and by caching and processing data in-network. However, without proposed geographic
routing support, there is initial and periodic interest and low
rate data flooding throughout the network. GEAR protocol
can compliment this work by efficiently route interest to the
destination region, thus conserve more energy.
Gao and Pottie [6, 5] proposes a table driven, multi-path
network structure for the communication between a large
number of sensors and a central information gathering entity called the USER. However, their pre-built routing table
and multi-path structure may not scale to large size sensor
networks. Moreover, table driven approach may not be able
to adapt well to network dynamics or traffic dynamics at a
low cost.
Localization work There has been substantial research interest [22, 18, 2, 7, 19] in localization systems. Such systems are a prerequisite for geographical routing and other
sensor net applications. Of those, Ward et al. [22] propose
a ultrasonic location system based on tri-lateration principle; Bulusu et al. [2] propose a coarse-grained connectivity
metric method for localization in outdoor environments in
the absence of GPS; Girod and Estrin [7] propose a robust
range estimation technique using acoustic and multimodal
sensing; Savvides et al. [19] propose Ad-Hoc Localization
system (AHLos), a find-grained localization technique for
ad-hoc sensor networks.
3 Geographical and Energy Aware Routing
(GEAR)
We now describe the Geographical and Energy Aware
Routing(GEAR) algorithm. As mentioned in the introduction, we are interested in routing queries to regions in proposed sensor-net applications. The process of forwarding a
packet to all the nodes in the target region consists of two
phases:
1. Forwarding the packets towards the target region:
GEAR uses a geographical and energy aware neighbor
selection heuristic to route the packet towards the target
region. There are two cases to consider:
(a) When a closer neighbor to the destination exists:
GEAR picks a next-hop node among all neighbors
that are closer to the destination.
(b) When all neighbors are further away: In this case,
there is a hole. GEAR picks a next-hop node that
minimizes some cost value of this neighbor, which
we will discuss in detail in section 3.1.2.
2. Disseminating the packet within the region:
Under most conditions, we use a Recursive Geographic
Forwarding algorithm to disseminate the packet within
the region. However, under some low density conditions, recursive geographic forwarding sometimes does
not terminate, routing uselessly around an empty target region before the packet’s hop-count exceeds some
bound. In these cases, we propose to use restricted
flooding.
Before we describe the above algorithms in detail, we
state the assumptions of this work:
1. each query packet has a target region specified in some
way (for the description of the algorithm, we assume a
rectangular region specification).
2. each node knows its own location and remaining energy level, and its neighbors’ locations and remaining
energy levels through a simple neighbor hello protocol. Note that a node can obtain its location information at low cost from GPS or some localization system [22, 18, 2, 7, 19], which presumably is already
available due to the needs of sensor net applications.
3. the link is bi-directional, i.e., if a node hears from
a neighbor N i , then its transmission range can reach
Ni . This is not an unreasonable choice as most MAC
layer protocols, such as IEEE 802.11, assume symmetric links.
3.1
Energy-Aware Neighbor Computation
In this section, we assume that the node N is forwarding
packet P , whose target region is R. The centroid of the target
region is D. Upon receiving a packet P , the node N routes
P progressively towards the target region, and at the same
time tries to balance the energy consumption across all its
neighbors. Node N achieves this trade-off by minimizing
the learned cost h(Ni ; R) value of its neighbor N i .
Each node N maintains state h(N; R) which we call its
learned cost to region R. A node infrequently updates its
h(N; R) value to its neighbors. We discuss this infrequent
update later. We implicitly define h(N; R) in the next two
paragraphs.
If a node does not have h(N i ; R) state for a neighbor N i ,
it computes the estimated cost c(N i ; R) as a default value
for h(Ni ; R). The estimated cost c(Ni ; R) of Ni is defined
as follows:
c(Ni ; R)
=
d(Ni ; R) + (1 )e(Ni )
(1)
where is a tunable weight, d(N i ; R) is the distance from
Ni to the centroid D of region R normalized by the largest
such distance among all neighbors of N , and e(N i ) is the
consumed energy at node N i normalized by the largest consumed energy among neighbors of N .
After a node picks a next-hop neighbor N min (we describe this neighbor selection in section 3.1.1 and section
3.1.2), it sets its own h(N; R) to h(N min ; R) + C (N; Nmin )
where the latter term is the cost of transmitting a packet from
N
to Nmin . C (N; Nmin ) can also be a combination function of both the remaining energy levels of N , N min and the
distance between these two neighbors.
The intuition for minimizing the estimated cost function
c(N; R) is as follows:
When all nodes have equal energy, this degenerates to
the classical greedy geographic forwarding: forwarding
the packet to the nearest neighbor to destination.
When all neighbors are equidistant, this degenerates to
load splitting among neighbors. Note that minimizing
the energy cost among neighbors is a local approximation to the lowest energy cost path, if it is more expensive to use a node that has less remaining energy. Since
GEAR makes forwarding decision only based on local
knowledge, an approximation to the global lowest cost
path is the best that a local algorithm can achieve.
Now that a node has a learned cost state h(N; R) or a
default estimated cost function c(N; R) for each neighbor,
we now describe the forwarding actions at node N . As with
other geographical routing schemes, there are two cases to
consider:
k
F
A
L
G
B
T
H
C
I
J
D
E
S
Figure 1: Learning routes around holes
Ni
If at least one neighbor of N is closer to D than N ;
All neighbors are further away from D than N .
3.1.1 Closer Neighbor Exists
As mentioned before, under GEAR, the packet contains a
target region field. Therefore, a forwarding node can make
locally greedy choice in selecting next-hop node. Whenever
a node N receives a packet, it will pick the next hop among
the neighbors that are closer to the destination, at the same
time, minimizing the learned cost value h(N i ; R). Since it
picks a next-hop node from closer neighbors, it will route
progressively towards the target region when there are no
holes. Without holes, the learned cost is a combination of
consumed energy and distance, minimizing the learned cost
value is a trade-off between routing towards the next-hop
closest to the destination and balancing energy usage.
3.1.2 All Neighbors are Farther Away
In this case, N knows it is in a hole. A node’s learned
cost h(N; R) and its update rule are combined to circumvent holes. Intuitively, when there is no hole in the path
towards R, the node’s learned cost h(N i ; R) is equivalent
to the estimated cost c(Ni ; R). However, when there is a
hole in the path towards R, the node’s learned cost represents “resistence” to following the path towards that hole;
“resistence” that the estimated cost cannot provide. Next,
we illustrate this latter feature of the learned cost.
Figure1 is a grid topology. Suppose the distance between
the nearest two neighbors is 1, and each node can reach its
8 neighbors. The nodes in black, i.e., G, H , I , are energy
depleted nodes, thereby can not relay packets. Suppose node
S wants to send a packet to region R with centroid at T . For
illustration purposes, we use T to denote this region. Again,
Figure 2: Recursive Geographic Forwarding
for simplicity, we illustrate the algorithm using pure geographic routing, i.e., we set in Equation 1 to be 1, and we
use direct distance instead of normalized distance mentioned
earlier.
Initially, at time 0, at node S , among all neighbors ofpS ,
B , C , D are closer to T than S . h(B; T ) = c(B;pT ) = 5,
h(C; T ) = c(C; T ) = 2, h(D; T ) = c(D; T ) = 5.
Upon receiving a packet destined to T , S will forward it
to its lowest cost neighbor, i.e., C . At C , it will find itself
in a hole, since all C ’s neighbors are further away from T
than itself. At C , it forwards the packet to the node with the
minimum h(N; T ). When there are ties, it breaks ties based
on some predefined ordering (e.g., node ID). For example,
it picks B as the next hop, then updatepits own h(C; T ) =
h(B; T ) + C (C; B ), where h(B; T ) = 5, and suppose one
hop transmission cost from a node to its neighbor is 1, i.e.,
C(C, B) = 1.
Later on, for example, at time 2, node S receives a packet
destined to the same region T , the h values of its neighbors
become:
p
p
p
h(B; T ) = 5, h(C; T ) = 5 + 1, h(D; T ) = 5.
Thus, this time node S will forward the packet directly to
B instead of C to circumvent the hole. The actual forwarding action of the algorithm at node S will oscillate between
B and C several rounds before the learned cost converges to
favoring B as a next hop neighbor. After the first packet
reaches the destination, the correct learned cost value will
be propagated one-hop back. Every time a packet is delivered, the correct learned cost value will be propagated onehop away. Therefore, suppose the path length from S to T
is n, the learned cost will converge after the node delivers
n packets to the same target T . Note that the convergence
of learned cost does not affect successfully routing a packet
out of holes, it only affects how efficient is the hole routing
path. Propagating the learned cost values further upstream
through the update rule will enable the packet to have an earlier chance to avoid holes (i.e., more effectively circumnavigate holes), and at the same time avoid depleting the nodes
surrounding the holes.
In summary, the learned cost together with its update rule
help to learn the route around holes. Intuitively, the learned
cost is set to the current best choice available.
This learned cost is inspired by the Learning Real Time
A* algorithm [14], which is a well known heuristic search
technique. In [14], Korf has proved the completeness of the
LRTA*, i.e., if there is a path, LRTA* will find it, and there
will not be infinite loop. Since our learned cost based hole
routing algorithm is very similar to LRTA*, we conjecture
that this result applies to our scheme as well.
3.1.3 Discussion
The estimated cost c(N; R) is a combination of the normalized distance from a neighbor to the destination and its
normalized remaining energy level. In equation 1, can be
adjusted to emphasize minimizing path length to the destination or balancing energy consumption. We tried several
variants of this estimated cost function. For example, we
tried different energy cost functions and different normalization denominators. The simulation results show that the
algorithm performance is not very sensitive to the particular estimated cost function. Our explanation is that it is the
comparison (relative value among all the neighbors), not the
absolute estimated cost value that matters, since it is used to
make a local selection among all the neighbors.
For computing c(N; R), each node needs to know neighbors’ energy levels and locations. A node also needs neighbors’ learned cost to make forwarding decisions. Various
techniques are possible: e.g., piggybacking these on data
traffic, requesting this information on demand, advertising
the information only when its value changes significantly, or
a combination of the above. In our simulation, we implement
a threshold advertising scheme.
We have verified through simulation (section 4.4) that the
threshold can be set in a manner that the overhead is insignificant without adversely affecting performance.
3.2
Recursive Geographic Forwarding
Before a packet reaches the target region R, we use the
forwarding rules described in the previous section. Once the
packet is inside the target region, a simple flooding with duplicate suppression scheme can be used to flood the packet
inside region R. However, flooding is expensive in terms
of energy consumption, due to the fact that in this simple
flooding scheme, every node has to broadcast once, and all
its neighbors receive this broadcast message. This is especially expensive in high-density networks, which is the
case for some proposed sensor net applications where nodes
are densely, and redundantly deployed for robustness. This
demonstrates the necessity to use an energy efficient routing
algorithm in place of flooding in disseminating the packet
inside the target region, or any kind of flooding in sensor
networks.
Therefore, we use a Recursive Geographic Forwarding
approach to disseminate the packet inside target region R.
As shown in Figure2, suppose the target region R is the big
rectangle in Figure2, and now node N i receives a packet P
for region R, and finds itself inside R. In this case, N i creates four new copies of P bound to 4 sub-regions(as shown
by 4 small rectangles in Figure2) of region R. Repeat this
recursive splitting and forwarding procedure until the stop
condition for recursive splitting and forwarding are satisfied.
The recursive splitting terminates if the current node is
the only one inside this sub-region. The criteria to determine
this is when the farthest point of the region is within a node’s
transmission range, but none of its neighbors are inside the
region. When no node is inside the sub-region, the packet is
dropped altogether. The rule to determine a sub-region being
empty is the same as the previous criteria.
We choose to divide region R into 4 subregions for convenience. Other ways of recursively visiting subregions are
also possible, though we do not consider them in this paper.
3.2.1 Pathologies
When network density is low recursive geographic forwarding is subject to two pathologies: inefficient transmissions and non-termination.
Inefficient transmissions In recursive geographic forwarding, in its recursive splitting process, to reach 4 subregions, a unicast packet is sent to its neighbors multiple
times, and it is received only by the intended receivers. In
contrast, restricted flooding exploits the broadcast medium
of the wireless channel, only sends one broadcast message
to all its neighbors, but every node in its transmission range
receives this broadcast message whether it is an intended receiver or not. Which approach is more energy efficient depends on the density of the target region. In low density scenarios, it is more energy efficient to use restricted flooding
than recursive geographic forwarding.
For example, in Figure3, the rectangle is the target region
R, there are 4 nodes, i.e., D, B, C, and E, in the target region,
and the transmission range of a node is about the distance
from D to B . Suppose the average number of neighbors of
a node is 4. The first node that the packet has reached in
region R, i.e., C , finds out that R is within its transmission
range (i.e., all other nodes in R are its neighbors). Then the
node can use one-hop flooding instead of sending 3 copies of
unicast message separately to the other 3 nodes in region R
as in recursive geographic forwarding case. Suppose sending
or receiving a packet consumes 1 unit of energy, then the best
that recursive geographic forwarding can achieve is to send 3
A
D
B
E
C
F
G
Figure 3: Recursive Geographic Forwarding vs. restricted flooding
K
E
H
B
A
C
F
L
Figure 4: Non-termination at empty region in low density case
unicast messages to the other 3 nodes in R, thus consuming
6 units of energy in total. However, one-hop flooding consumes 1 unit energy for sending, and 4 units of energy for
receiving at C ’s neighbors (i.e., B, D, E, F), thus 5 units of
energy in total. This is just a simple example to show the
difference between two schemes. Extrapolated to a larger
region, the differences can be significant.
Non-termination In the recursive geographic forwarding
protocol, packet forwarding terminates when the target subregion is empty. The following heuristic is used to determine
if the target sub-region is empty: if the farthest point of the
region is within its transmission range, but none of its neighbors are inside the region, the region is considered empty.
However, this strategy does not work when the network
density is low compared to the (sub)target region size. When
density is low the probability of the target region being
empty is high, and the transmission range is small compared
to the target region size so that the node that is closest to the
target region can not reach the other end of the target region.
As a result, the nodes outside the target region have no indication that the region is empty, hence the packet still searches
for routes to get into the empty region. The search will not
stop until the number of hops traversed exceeds the packet’s
time-to-live. This failed search process can heavily drain the
nodes around the target region.
For example, in Figure4, suppose the blue (gray) rectangle is one subregion of the target region R, namely R s ,
toward which the packet is heading. Each node has equal
transmission range, and the dotted circle indicates the transmission range of C . In this example, region R s is empty,
however, for each individual node bordering R s , i.e., C, E,
K, L, its transmission range is unable to cover the whole region Rs . For example, at C , its transmission range can not
reach the far end of region R s (i.e., H ). Therefore, the packet
can not tell that region R s is empty. Hence, the packet will
repeatedly traverse the nodes around region R s until its hop
count passes some bound.
We propose to use restricted flooding to deal with these
pathologies in low density scenarios. Node degree is used
as a criteria to differentiate low density scenarios from high
density scenarios. Once the packet reaches the first node J
in the target region, whether to use recursive geographic forwarding or restricted flooding will be based on the number
of neighbors of node J . If this number is below a threshold,
then the packet is flooded inside the region, otherwise recursive geographic forwarding will be triggered. Simulation results show that this approach helps to avoid the pathological
case mentioned before, thereby improving the performance
dramatically(more than an order of magnitude improvement)
in low density scenarios, and exhibits the same performance
as before in high density case, since restricted flooding is not
triggered in high density scenarios. However, we will explore a better alternative solution to this problem as future
work.
3.3
Adaptive energy aware behavior of our algorithm
We use an energy aware metric in the estimated cost function to balance energy consumption, and achieve energy efficiency for the whole network. However, under some circumstances, taking the geographically direct path is more energy
efficient, and consequently prolongs network lifetime. Under the following conditions, the pure geographic mode is
used instead of adding the energy aware metric.
1. After the number of hops traversed crosses some threshold, pure geographical routing is used instead of energy
aware routing. The motivation behind this is to get to
the target directly if the packet has already traveled a
long way. The threshold is tunable, however, the algorithm is not very sensitive to this parameter which is
confirmed by the simulation results. For example, it can
be set to a large enough number.
2. After a packet reaches a node whose neighbors are
heavily depleted (which indicates a neighborhood
where nodes are heavily depleted), the packet will
switch to pure geographic mode to avoid taking an alternative longer path, and consequently consuming more
energy than the direct path.
3. When nodes are near the target region, pure geographic
mode is used. Since the packet is supposed to be disseminated to all the nodes in the target region, the nodes
bordering the target region are the bottleneck. Pure geographic mode is used near the target region to avoid
taking a longer path and thereby avoid burning out more
bottleneck nodes more quickly.
idle state will dominate the total energy consumption
of a node, and make different routing protocols exhibit
roughly the same energy consumption characteristics.
As a result, sensor net applications demand a MAC
layer where a node can put itself into sleep mode when
it is not involved in active communication [20]. On the
other hand, our routing protocol does not depend on any
particular type of MAC protocol as long as it can resolve medium access control issue, therefore, we use a
MAC layer that assumes no collision. Note that our
goal is performance comparison between different protocols. The effect of this idealization is common to all
simulated protocols, therefore, the relative comparison
between different protocols should not change much
under a different MAC layer, even though a MAC layer
with more realistic details may change a routing protocol’s performance.
4 Simulation and comparison
Geographic routing is designed to improve efficiency relative to flooding data. Our primary performance measure is
increased network lifetime due to energy awareness. Moreover, we are interested in how this comparative measure
scales with network size. Sensor networks may involve thousands of nodes, thus scaling to large size network is essential
for a routing protocol to be applicable to sensor networks.
We study and compare GEAR with and without energy
awareness. In addition, we evaluate how our geographic
routing algorithm compares with other geographic routing
algorithm, such as GPSR where a pro-active and deterministic hole routing scheme is used. Since the original GPSR
does not handle routing to a region, we have augmented
GPSR with our recursive geographic forwarding protocol to
route packets to a region. Idealized multicast and flooding
are simulated for comparison. They represent two extremes,
idealized multicast requires global information to compute
shortest paths to every node in the target region, while flooding uses no information, but is far less efficient.
Testing the scalability of our protocol to large networks
was our primary goal, and we achieved this through simulation. We are also completing a prototype implementation to
validate our results on moderate size testbed.
Next, we describe and justify some major features of our
simulator.
We use a discrete event-driven simulator. It simulates
routing packets among different nodes in the network.
We are not interested in the processing details of each
packet in the network per se, therefore we chose to leave
out packet level details, such as transmission delay, or
queuing delay etc. Leaving out irrelevant details enables us to simulate large size network.
We use a simple energy model in which every node
starts with the same initial energy budget, and transmitting or receiving a packet consumes one unit of energy.
We do not count energy consumed by control packets.
As we show in section 4.4, control overhead is relatively small in GEAR. Therefore, our comparison is
fair considering that periodic updating in other protocols may consume much more energy than threshold
triggered update in our approach.
MAC assumptions: We assume a MAC layer in which
each node consumes negligible energy when it is not
sending or receiving packets. In a MAC layer which
can not turn itself off when it is not actively transmitting or receiving, the energy consumed by listening or
We adjust the following parameters in the simulation: network size, density, size of target region, and traffic description.
In this paper, we focus on varying network size, keeping
density and other parameters constant. Our goal is to design a protocol that can scale to thousands of nodes in future
sensor networks, therefore we focus our study on how the
protocol scales with networks of different sizes. However,
we also conducted some scalability studies with other input
parameters. The simulation results show that the GEAR algorithm also scales well along other input parameters.
4.1
Simulation Scenarios
We do not know how future sensor net traffic will be, thus
we choose two simple models: uniform and non-uniform
traffic.
Uniform Traffic Randomly distributed source and destination pairs. The source and target regions are randomly
distributed throughout the network. This represents independently and uniformly distributed traffic.
Non-uniform Clustered traffic For clustered source and
destination pairs, the sources and destinations are randomly selected, but the sources from different traffic
pairs, or the destinations from different traffic pairs are
geographically close to each other. This represents nonuniform traffic distribution, which is not uncommon in
proposed sensor net applications where detected physical events or triggered communication are not completely independent from each other, and sometimes
highly correlated.
As mentioned earlier, in this paper, we vary network size,
fix density, transmission range, and other parameters. More
specifically, the simulation results shown here include networks of size ranging from 400, to 4800 nodes. For a 600
nodes network, its geometric area is 1200 x 1200 square. We
adjust network geometric area proportionally to the number
of nodes in the network to keep network density fixed. We fix
a node’s transmission range to be 100 units across all simulations. The target region is a circle with radius 50 units.
Performance Metrics
It is difficult to precisely define metrics to study the impact of energy awareness on network lifetime. However, we
have developed two metrics that approximately capture notions of network lifetime:
Packets before partition Number of data packets sent and
successfully delivered before network partition. Network is considered partitioned if all the given sources
are partitioned from their respective target regions. In
some sense, this metric indicates the network lifetime.
Connectivity after partition Fraction of pairs still connected after partition. This metric indicates how the
given traffic affects the rest of the network. Because we
lack of benchmark traffic model, we concern the residual energy after the given traffic pairs are partitioned,
and how much further communication it can support.
4.3
Simulation Results
Uniform Traffic With random/uniform traffic distribution,
our algorithm successfully delivers 0.25 to 0.35 times more
packets than GPSR. GEAR is also more energy efficient,
which is measured in terms of the “connectivity after partition” metric mentioned earlier. In both uniform traffic distribution and non-uniform traffic distribution (which is shown
in Figure 8), GEAR delivers significantly (40 to 100 times)
more packets than flooding.
Figure 5 shows the mean and 95% confidence interval of
the number of packets sent and successfully delivered before
network partition. Note that some confidence interval is too
tight to be visible in the graph. All later graphs show the
mean and 95% confidence interval of the y-axis value, except Figure 9 and Figure10, where we omit the confidence
interval to make the graph easier to read.
As shown in Figure 5, GEAR successfully delivers 25 35% more packets than GPSR. The energy aware version
delivers comparable number of packets compared to its pure
geographical variant. In complete random/ uniform traffic
distribution, the energy consumption is already balanced due
to randomly distributed traffic. Extra energy balancing efforts will not help much in terms of contributing to delivering more packets before network partition. Therefore, under
uniform traffic, the energy aware version and the pure geographical version deliver similar number of packets before
network partition. The reason that GEAR prolongs network
lifetime compared to GPSR, is that GPSR tends to concentrate traffic on the perimeter when it routes around holes,
thus burning out the nodes on the perimeter sooner. This is
confirmed by the fact that under GPSR, after network partition, there are more nodes depleted along the direct paths
Number of packets Sent(or unique pkts received) before network partition
4.2
35000
GPSR + RGF
Flooding
Idealized multicast
GEAR(RGF + pure GR)
GEAR(RGF + Energy Aware)
30000
25000
20000
15000
10000
5000
0
400600800
1200
2400
Network size: Number of nodes in the system
4800
Figure 5: Comparison for uniform traffic
1.4
Fraction of pairs remain connected after network partition
10 traffic source and target pairs are randomly selected either from a uniform distribution or clustered together. A
node’s initial energy level is 1 joule, transmiting or receiving a packet consumes 0.001 joule. Each data point in the
graphs presented in this paper is averaged over at least 100
simulation runs. For the results presented in this section, we
set in Equation 1 to 0:5.
GPSR + RGF
Flooding
Idealized multicast
GEAR: RGF + pure GR
GEAR: RGF + Energy Aware
1.2
1
0.8
0.6
0.4
0.2
0
400 600 800
1200
2400
Network size: Number of nodes in the system
4800
Figure 6: Fraction of pairs remain connected after partition
for uniform traffic
between the source and the target compared to the GEAR
case. As expected, the idealized multicast protocol delivers
the largest number of packets before network partition in this
uniform traffic case since we did not count energy consumption for control packets in the simulation. However, in reality, it is prohibitively expensive to collect the required global
information for idealized multicast routing.
As shown in Figure 6, after the given traffic pairs are partitioned, there are more source and destination pairs remaining
connected in GEAR than other protocols. This is an energyefficient feature in the sense that after the given traffic pairs
are partitioned, the rest of the network can still communicate.
However, this metric alone can not represent the energy efficiency of a protocol, since an inefficient protocol may deliver far fewer packets than other protocols before network
partition, thereby having lots of energy left in the network,
and consequently having more pairs connected after partition. For example, as shown in Figure 5, GPSR delivers between 25% and 35% fewer packets before network partition
than GEAR, but its curve is very close to GEAR’s in Figure 6. To better differentiate between these two protocols
in terms of their energy efficiency, we introduce the metric
resource expended per packet delivered, and we count resources expended in terms of the number of connected pairs
that are broken down because of nodes being depleted. Precisely, it is defined as:
Nb Ne
total number of delivered packets
(2)
where Nb is total number of connected pairs in the beginning, Ne is the is total number of connected pairs after network partition. It can be considered as a normalized version
of the metric presented in Figure 6.
GPSR + RGF
Idealized multicast
GEAR: RGF + pure GR
GEAR: RGF + Energy Aware
Number of connected pairs torn down per delivered packet
140
120
100
80
60
40
20
0
400 600 800
1200
2400
4800
Network size: Number of nodes in the system
Number of packets Sent(or unique pkts received) before network partition
Figure 7: Number of broken down pairs per delivered data
packet
8000
GPSR + RGF
Flooding
Idealized multicast
GEAR(RGF + pure GR)
GEAR(RGF + Energy Aware)
7000
6000
5000
4000
3000
2000
1000
0
400 600
1200
1600 2000 2400
Network size: Number of nodes in the system
4800
Figure 8: Comparison for non-uniform traffic
Figure 7 shows the number of pairs disconnected per delivered packet with different protocols. Note that the lower
is the curve, the more energy efficient is the protocol. As
shown in the graph, GEAR is much better than all other protocols, including the pure geographical variant of the same
algorithm. Specifically, GEAR is approximately 50% more
efficient than its pure geographic variant, approximately 70%
more efficient than GPSR, and approximately 80% more efficient than idealized multicast. The value for Flooding is
too large to be shown in this graph. This may suggest that
the GEAR protocol successfully achieves energy efficiency.
Note that the metric in Figure 7 magnifies the difference between different protocols in terms of energy efficiency compared to the metric in Figure 6.
Non-uniform Traffic Clustered source and destination
pairs represent non-uniform traffic distribution in sensor net
applications.
Figure 8 shows that GEAR delivers 70 - 80% more packets than GPSR and 4 - 15% fewer packets than idealized multicast. Hence, GEAR exhibits more gain in non-uniform traffic scenarios than uniform traffic scenarios. Our explanation
is that when traffic sources are clustered together, GEAR’s
energy balancing efforts pay off most.
Under non-uniform energy distribution(i.e., the nodes in
the target region have higher initial energy level than others), the gains for GEAR are even much higher. Simulation
results show that GEAR delivers 125% more packets than
GPSR and 25% more packets than idealized multicast. That
GEAR performs better than idealized multicast seems surprising. However, our energy aware routing scheme balances
energy consumption, while the idealized multicast keeps using the shortest path, which in turn will burn out the nodes
along those paths quickly. Note that non-uniform energy distribution may map to real world sensor networks where some
nodes that have sensing capability may have different energy
resource than others.
4.4
Trade-off between neighbor information update and performance
The results shown in this section are for 600 nodes network, traffic are uniformly distributed, a node’s initial energy
level is 1 joule, transmitting or receiving a packet consumes
0.001 joule. As mentioned before, neighbor information update can be triggered by a pre-defined threshold. As shown in
Figure 9 and Figure 10, it is not necessary to update neighbor information for every packet. With increasing update
threshold, the protocol performance degrades gracefully, but
the number of control packets generated drops dramatically.
For example, in Figure 10, for the point with learned cost
h-value update threshold = 5, and remaining energy update
threshold = 0.05, the average number of control packets sent
per node for the whole simulation run is only 32 packets, i.e.,
3.2% of a node’s total energy budget. However, as shown in
Figrure 9, the performance for this particular case only degrades approximately 6% compared to the no update delay
case, which corresponds to the black square in Figrure 9.
5 Discussion and Future work
Sensitivity to location error: We also studied how sensitive the protocol is to the location error. In the real world,
location error can be caused by either imprecise measurement from GPS or localization system, or nodes moving, but
failing to update location information timely. To investigate
the sensitivity of our scheme to location error, we introduce
a random location error for every node in the network. Simulation results show that with moderate location error, the
Number of Delivered packets before network partition
22000
no update delay
discrete remaining energy update interval: 0.05
discrete remaining energy update interval: 0.1
discrete remaining energy update interval: 0.4
discrete remaining energy update interval: 1
21000
20000
19000
18000
17000
16000
15000
0.5
1
5
Learned cost(h-value) update interval
10
Average number of control packets sent per node for the whole simulation run
Figure 9: Number of delivered packets with different neighbor
information update threshold
120
Discrete remaining energy update threshold: 0.05
Discrete remaining energy update threshold: 0.1
Discrete remaining energy update threshold: 0.4
Discrete remaining energy update threshold: 1
100
80
60
40
20
0
0.5
1
5
Learned cost(h_value) update threshold
10
Figure 10: Number of update control packets with different
neighbor information update threshold
protocol still achieves satisfying performance. For example, for 1200 x 1200 square, and 600 nodes network, each
node’s transmission range is 100 units, when we introduce
random error in the range [-6, 6] for each node’s x- and ycoordinates, the simulation results show no statistical performance degradation. If we introduce randomized error [-60,
60] for each node’s x- and y- coordinates, the simulation results show only 9% performance degradation.
Impact on average path length: GEAR achieves energy
balancing by taking alternative path, therefore, it is not surprising that our energy balancing strategy increases path
length by 25% to 45% over all packets delivered. However,
GEAR delivers more packets than GPSR. When we check
the initial same number of delivered packets as in GPSR, the
average path length for those packets in GEAR is only 10%
longer than GPSR. After some nodes along the path are depleted, the packet has to take alternative longer path to avoid
holes, thus the average path length for later communication
is expected to be longer than those earlier on.
Implementation and Porting to ns-2 We have already
implemented a pure geographic variant of GEAR protocol
using our data dissemination diffusion testbed [10]. As a
next step, we plan to implement the full-fledged version of
GEAR. We plan to investigate how the details that our simulator abstracted out affect the protocol performance, specially how the real world MAC affects the protocol performance in the implementation. We also plan to study how the
GEAR algorithm interacts with the directed diffusion implementation.
We also plan to port the GEAR algorithm to ns-2 [1], so
that people can study how geographical and energy aware
routing interacts with other protocols, such as directed diffusion in sensor networks, with the presence of a more detailed
lower layer but still in a controlled simulation environment.
6 Conclusion
Motivated by future sensor network applications, we studied the problem of forwarding a packet to nodes in a geographic region of an ad-hoc wireless sensor network. The
proposed Geographic and Energy Aware Routing (GEAR)
protocol uses energy aware and geographically informed
neighbor selection to route a packet towards the target region. This strategy attempts to balance energy consumption
and thereby increase network lifetime. Within a region, it
uses a recursive geographic forwarding technique to disseminate the packet. The simulation results show that for an uneven traffic distributions, GEAR delivers 70% to 80% more
packets than GPSR. For uniform traffic pairs, GEAR delivers 25 - 35% more packets than GPSR. Moreover, in both
cases, GEAR performs better in terms of connectivity after
initial partition. We are currently implementing a prototype
of the GEAR protocol in a moderate size testbed. We plan
to investigate how the details of a real implementation affect
the protocol performance.
References
[1] Sandeep Bajaj, Lee Breslau, Deborah Estrin, Kevin
Fall, Sally Floyd, Padma Haldar, Mark Handley,
Ahmed Helmy, John Heidemann, Polly Huang, Satish
Kumar, Steven McCanne, Reza Rejaie, Puneet Sharma,
Kannan Varadhan, Ya Xu, Haobo Yu, and Daniel Zappala. Improving simulation for network research. Technical Report 99-702b, University of Southern California, March 1999. revised September 1999, to appear in
IEEE Computer.
[2] Nirupama Bulusu, John Heidemann, and Deborah Estrin. Gps-less low cost outdoor localization for very
small devices. IEEE Personal Communications Magazine, 7(5):28–34, October 2000.
[3] Jae-Hwan Chang and Leandros Tassiulas. Energy conserving routing in wireless ad-hoc networks. INFOCOM, 2000.
[4] Gregory G. Finn. Routing and addressing problems in
large metropolitan-scale internetworks. Technical Report ISI/RR-87-180, USC/ISI, March 1987.
[18] Nissanka B. Priyantha, Anit Chakraborty, and Hari Balakrishnan. The cricket location-support system. In
Proc. ACM Mobicom, Boston, MA, 2000.
[5] J. Gao. Energy efficient routing for wireless sensor networks. Ph.D. Dissertation, UCLA, 2000.
[19] Andreas Savvides, Chih-Chieh Han, and Mani B.
Strivastava. Dynamic fine-grained localization in adhoc networks of sensors. In Proc. ACM Mobicom,
2001.
[6] J. Gao, K. Sohrabi, V. Ailawadhi, and G. Pottie. An
energy conscious self-organizing wireless sensor network. IEEE PCI Magazine, October 2000.
[7] L. Girod and D. Estrin. Robust range estimation using
acoustic and multimodal sensing. In IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS 2001), Maui, Hawaii, October 2001.
[8] W. Rabiner Heinzelman, A. Chandrakasan, and Hari
Balakrishnan. Energy-Efficient Communication Protocol for Wireless Microsensor Networks. In Proceedings of the 33rd International Conference on System
Sciences (HICSS ’00), 2000.
[9] Tomasz Imielinski and Samir Goel. Dataspace - querying and monitoring deeply networked collections in
physical space. IEEE Personal Communications Magazine, August 2000.
[10] Chalermek Intanagonwiwat, Ramesh Govindan, and
Deborah Estrin. Directed diffusion: A scalable and robust communication paradigm for sensor networks. In
Proc. ACM Mobicom, Boston MA, 2000.
[11] D. B. Johnson and D. A. Maltz. Dynamic Source Routing in Ad-hoc Wireless Networks. In T. Imielinksi and
H. Korth, editors, Mobile Computing, pages 153–181.
Kluwer Academic Publishers, 1996.
[12] Brad Karp and H. T. Kung. Gpsr: Greedy perimeter
stateless routing for wireless networks. In Proc. ACM
Mobicom, Boston, MA, 2000.
[13] Yong-Bae Ko and Nitin H. Vaidya. Location-Aided
Routing (LAR) in Mobile Ad Hoc Networks. In Proceedings of the Fourth Annual ACM/IEEE International Conference on Mobile Computing and Networking (Mobicom’98), Dallas, TX, 1998.
[14] Richard Korf. Real-time heuristic search. Artificial Intelligence, 42(2-3):189–211, March 1990.
[15] Jinyang Li, John Jannotti, Douglas S. J. De Couto,
David R. Karger, and Robert Morris. A scalable location service for geographic ad hoc routing. In Proc.
ACM Mobicom, Boston, MA, 2000.
[16] V. D. Park and M. S. Corson. A Highly Adaptive Distributed Routing Algorithm for Mobile Wireless Networks. In Proceedings of INFOCOM 97, pages 1405–
1413, April 1997.
[17] Charles Perkins. Ad-Hoc On Demand Distance Vector Routing (AODV). Internet-Draft, November 1997.
draft-ietf-manet-aodv-00.txt.
[20] S. Singh and C.S. Raghavendra. Pamas: Power aware
multi-access protocol with signalling for ad hoc networks. ACM CCR, July 1998.
[21] Mark Stemm and Randy H. Katz. Measuring and reducing energy consumption of network interfaces in
hand-held devices. IEICE (Institute of Electronics,
Information and Communication Engineers) Transactions on Communications, Special Issue on Mobile
Computing.
[22] Andy Ward, Alan Jones, and A. Hopper. A new location technique for the active office. IEEE Personal
Communications, 4(5), October 1997.
[23] Mike Woo, Suresh Singh, and C. S. Raghavendra.
Power-aware routing in mobile ad hoc networks.
ACM/IEEE MOBICOM, 1998.
[24] Seung-Chul Woo and Suresh Singh. Scalable routing
protocol for ad hoc networks. accepted for publication
in Journal of Wireless Networks (WINET).